Native C++ FFI in the same manner as D

[bstrie]

@sanxiyn and @jdm have expressed interest in doing this, possibly in the 0.8 cycle.

http://www.reddit.com/r/rust/comments/1c3clf/c_ffi/c9codm1

[bstrie]

Here's a resource describing the level to which D can interface with C++:

http://dlang.org/cpp_interface.html

[asb]

It's worth taking a look at what the Julia community have been working on with C++ support via Clang.jl JuliaInterop/Clang.jl#20

[emberian]

Still very relevant

[graydon]

just a bug, removing milestone/nomination.

[metajack]

Visiting for triage. Nothing to add.

[klutzy]

Taking examples from D FFI doc,

int foo(int i, int j, int k);

can be expressed in Rust as

extern "C++" {
    fn foo(i: c_int, j: c_int, k: c_int) -> c_int;
}

This is straightforward; I've locally tuned rustc with tiny proof-of-concept mangler and it worked successfully.

However, if C++ code is in some namespace e.g.

namespace A {
    namespace B {
        int foo(int i, int j, int k);
    }
}

then we need to specify namespace in Rust since mangled function name contains it (at least on linux). One way to achieve that is:

#[namespace = "A::B"]
extern "C++" {
    fn foo(i: c_int, j: c_int, k: c_int) -> c_int;
}

I didn't tried to implement it, but it is still not hard; the only stuff I have to do is name mangling.

Classes and methods are more challenging: we need discussion on syntax. Again from D example,

class D {
public:
   virtual int bar(int i, int j, int k);
}
D *getD();

How this can be expressed in Rust? I've initially thought:

extern "C++" {
    trait D {
        fn bar(i: c_int, j: c_int, k: c_int) -> c_int;
    }
}

but D* getD(); becomes tricky since we don't have raw pointer to traits. Maybe we can treat extern C++ trait as structs as well, but it may be unintuitive for users.

[pnkfelix]

1. Could this be a motivation for adding D* (where D is a trait) to our grammar, even if its only to support a detail of the FFI? (In particular, I imagine you would not be able to do D* as D& etc...)
   Ugh, even just writing it down makes me say "no that seems bad."
2. We don't have to shoehorn Rust's syntax into something matching C++ semantics, because we have macros, right?
   (Obviously the macros will need to then expand into something, and so its important to specify what that something is, in terms of what the interface to the end user is. E.g. as @klutzy points out, we still need to figure out what class D maps to. And of course we don't have macro-expansion in all the relevant contexts, such as type-expressions, so you'd still want the macro to generate bindings for sane identifiers that can be used elsewhere directly. But I think we can separate the two concerns here...)
3. Ignoring the question of macros and continuing with @klutzy's idea of leveraging attributes, we could map each C++ class to two distinct bindings in Rust: one for its data layout (--> struct), and one for its vtable (--> trait).

For example:

class S {
  int x;
  virtual void foo();
  void bar();
}
class T : public S {
  int y;
  void baz();
}
void process(S* arg);
S* getS();

could be expressed as:

extern "C++" {
  #[class = "S"]
  struct Simpl { x: int, }

  #[class = "S"]
  trait Stable {
    fn foo(*self),
  }
  impl Simpl { fn bar(*self); }
  impl Stable for Simpl { fn foo(*self); }

  #[class = "T"]
  struct Timpl { super_: Simpl, y: int, }
  impl Stable for Timpl { fn foo(*self); }
  impl Timpl { fn baz(*self); }

  fn process<S:Stable>(arg: *S);
  fn getS() -> *Simpl;
}

(one could imagine rearranging the above a bit, pulling the struct items out of the extern block, and moving their #[class=..] attributes down instead to the impl Simpl and impl Timpl items. Not sure if that helps or hurts. Anyway I imagine the trait item needs to stay in the extern block, since a C++ vtable need not be compatible with a Rust trait vtable.)

[klutzy]

On extern "C++" I wrote above: I forgot the Windows situation! We also have to specify calling convention on Win32 since mangling uses callconv info. See also mangling scheme by @lifthrasiir.

Maybe extern "C++" is wrongly placed since "C++" is not a calling convention as we used. So how to declare C++-mangled stdcall function? #[cxx] extern "stdcall"? extern "C++" "stdcall"?

[flaper87]

visiting for triage. Nothing to add here.

[klutzy]

I started to think we don't need language-level support for C++ FFI. For name mangling we can just use #[link_name = "_Z5func2v"]. For virtual methods, we can define
virtual method table as struct.
What we really need is tool for generating C++ ffi, like rust-bindgen. For COM, we need code generator for MIDL.

[nathansobo]

I would love to use Rust to write V8 extensions, and a direct C++ interface could make that a good experience.

[emberian]

Nominating for P-high. I imagine this is going to be pretty important for driving adoption.